A reading path through thirty years of claim and rebuttal
This page exists to answer one question with evidence: has anyone seriously studied the poised state?
It is not a survey. It is a guided reading path. For each work you get an overview first — what it claimed, what happened to it, what to watch for — and only then the pointer to the source. Nothing below assumes you have read any of these papers. If you want the shortest route, skip to the reading path; everything before it exists to make that list make sense.
The short answer, stated once so the rest can be read against it. Three different things are called “poised”. They have three different reception histories. One was tested and largely deflated; one is a large, live, genuinely contested field; one was neither tested nor refuted. Each component has a serious literature. The conjunction of all three has none, under any name.
The companion page, Shadows of the Poised State, presents the affirmative side: twelve systems scored against a five-feature framework. This page is the literature that challenges the premise of that scorecard. Read them together, in either order.
Almost every confusion in this literature comes from one word doing three jobs. The three senses share the English word “poised” and very little else — not an order parameter, not an observable, not a test. Before anything else, separate them.
The oldest sense. A network of interacting on/off elements — genes, say — can be in an ordered regime, where a perturbation dies out and the system is rigid, or a chaotic regime, where a perturbation spreads and the system is useless for storing anything. The claim is that useful systems sit at the boundary. The lineage runs from Kauffman’s randomly-wired macromolecular systems1 through autocatalytic sets of proteins2 to the book-length synthesis, The Origins of Order.3
A cellular-automata wing was bolted on at the end of the 1980s: Langton supplied a scalar control parameter λ and the claim that computational capacity is maximised near the ordered/chaotic boundary,4 and Packard reported an evolutionary experiment that appeared to confirm it.5 That wing is Act I, and it did not survive. The autocatalytic-set wing is Act IV, and it did.
A different intellectual object with a different pedigree. Here you record many units at once — neurons on an electrode array, birds in a flock, positions in a protein family — estimate the joint probability distribution over network states, and ask whether the model you fit sits at a special point in its own parameter space. The anchors are Beggs & Plenz’s neuronal avalanches,6 Mora & Bialek’s review,7 and Muñoz’s Reviews of Modern Physics Colloquium.8
This is largely not Kauffman’s claim, and it does not present itself as such. Its machinery — maximum-entropy models, a density of states, a fictitious temperature — has no Boolean-network content. In Muñoz’s review Kauffman is cited specifically for gene regulatory networks, as one branch among several, not as the source of the general hypothesis. This is Act II and Act III, and it is unsettled.
The narrowest and most recent sense: the proposal that the electronic Hamiltonians of biomolecules sit at the critical point of the Anderson metal–insulator transitionA disordered solid conducts when its electron states are spread out (metallic) and insulates when they are trapped on individual sites (localized). Turning up the disorder drives a transition between the two. At the transition itself the states are neither — they are multifractal, spread out but ragged., where the statistics of neighbouring energy levels are semi-Poisson rather than Poisson (insulating) or Wigner–Dyson (metallic), so that quantum coherence survives long enough to matter while charge transport remains possible. The flagship is Vattay, Kauffman & Niiranen 2014;9 the empirical companion is Vattay et al. 2015.10 This is Act V, and it is the most interesting result on this page.
Because the three senses have opposite outcomes, they cannot be cited together as mutually supporting. Sense (a) was deflated in one wing and formalised in another; sense (b) is contested; sense (c) is untested. Any argument that treats “the poised state” as one well-supported idea is silently averaging a refutation, a live dispute and a blank.
A one-dimensional cellular automatonA line of cells, each holding 0 or 1. Every cell updates simultaneously according to a lookup table that maps the pattern in its neighbourhood to its next value. The lookup table is the ‘rule’. is defined by a lookup table. Langton’s λ is simply the fraction of entries in that table whose output is not the quiescent state — for a two-state automaton with 0 quiescent, the fraction of 1s in the table. Sweep λ from 0 upward and average behaviour is claimed to run fixed point → periodic → “complex” → chaotic, with the interesting regime at a critical value λc.4
Packard then ran a genetic algorithm that evolved automata to perform a genuinely non-trivial computation — decide whether the initial line contains more 1s than 0s, a task requiring information to travel the whole lattice — and plotted a histogram of the evolved rules against λ. The histogram was bimodal, with clusters near λ ≈ 0.23 and λ ≈ 0.83, straddling the two critical regions. It was read as evidence that evolution selects the edge of chaos.5
Mitchell, Hraber & Crutchfield re-ran the experiment carefully.11 Their evolved rules clustered at λ ≈ 0.43 and λ ≈ 0.57 — four times closer to 1/2 than Packard’s peaks — and the critical-λ bins were empty. They then identified three mechanisms, none of which requires any critical dynamics, that together produce the two-peaks-around-a-dip shape.
The density task is invariant under exchanging 0 and 1. Suppose a competent rule had λ < 1/2: more neighbourhoods map to 0 than to 1, so on some inputs that genuinely contain a majority of 1s the rule reduces the count of 1s and misclassifies. The mirror argument bounds it from above. So any rule that performs the task well must have λ ≈ 1/2 a priori — as a property of the task’s symmetry, not of the rule’s dynamics. This is a prediction that could have been made without running anything, and it is incompatible with the hypothesis that selection should drive λ to 0.23 or 0.83.
The number of rules at a given λ is a binomial coefficient, |CA(λ,n)| = C(2n, λ2n), overwhelmingly concentrated at λ = 1/2 with variance ∝ 2−n. The paper’s control experiment runs the identical genetic algorithm with fitness never evaluated and the surviving population chosen at random. It collapses onto λ = 1/2 by generation 10. With selection switched on, the population is indistinguishable from that null until generation 35.
This is the cleanest worked null model in the whole poised-state literature, and it costs one extra run.
Around generation 20 each run commits to one of two mirror strategies: grow blocks of 1s (which needs a low-λ safety margin, so the rule does not manufacture a block from a minority input) or grow blocks of 0s (high λ). Individual runs land in one lobe only — 14 of 30 low, 16 of 30 high. The bimodal histogram is an artifact of pooling runs, not a bimodality within any population.
Two further findings matter. The evolved rules over-fit the lattice size: the best rule by strict performance (λ ≈ 0.54) scores about 0.95 on the 149-cell lattice it was evolved on and about 0.05 at 999 cells — it had learned the lattice, not the computation. And the evidential gap in the original was procedural: per-bin fitness was never reported, so nobody could tell whether the critical-λ bins held good rules or merely many rules.
What was demonstrated to be an artifact is the genetic-algorithm and statistical methodology of one published experiment — not a property of cellular automata, and not the general hypothesis. The authors say so themselves: their results “do not disprove the hypothesis that computational capability can be correlated with phase transitions in CA rule space… We have shown only that the published experimental support… was not reproduced.” They also note the discrepancy with the original is not fully explained; the leading candidate — extra injected randomness and a much higher mutation rate in the original setup — rests on a personal communication and was not re-implemented.
That is the honest statement. It is also, in practice, decisive: the strong form did not survive. The conceptual companion paper makes the same case against the assumed link between dynamical class, λ, and computational capability.12
Modern re-tests deflate the claim further rather than restoring it. In spiking microcircuits the edge predicts good parameter regions but is neither necessary nor sufficient.13 On neuromorphic hardware, critical dynamics help only for tasks demanding long memory and integration, and are actively harmful otherwise.14 Computational capacity can decrease on approach to the edge.15 Reservoir-computing performance turns out to be governed by warm-up and nonlinearity matching rather than proximity to a critical regime.16 The one substantial revival is in a different substrate — recurrent threshold-gate networks, where a computable critical line does track real-time computational performance.17
Net: the claim survives as a substrate-specific heuristic. It does not survive as a law, and the specific experiment that launched it did not replicate. That the field is still publishing surveys titled “Revisiting the edge of chaos: Again?” thirty-plus years later18 tells you something about how unevenly the deflation propagated.
This is the strongest version of the poised-state claim, and it deserves to be understood before it is criticised. Two signatures do most of the work. Both are stated below in plain language first.
Record N units, chop time into bins, and write down the on/off pattern in each bin. That pattern is a state. Now count how often each state occurs, sort states from most to least frequent, and plot frequency against rank on log–log axes. If the plot is a straight line of slope −1 — the second-commonest state half as common as the commonest, the hundredth a hundredth as common — that is Zipf’s law.
Mora & Bialek turn this into a thermodynamic statement in three steps.7 First, define an energy from the observed probability, with the temperature fixed at 1 by convention:
Second, note that the rank of a state is exactly the number of states with lower energy, so log(rank) is the entropy at that energy. Third, substitute: Zipf’s law becomes the statement that entropy is a perfectly linear function of energy,
Why that is “critical”: reintroduce a fictitious temperature and the usual saddle point puts the system where dS/dε = 1/kBT. Ordinarily that picks out one temperature. Under Zipf’s law dS/dε equals 1/α at every energy, so kBT = α is a critical point — and because the system operates at kBT = 1, a Zipf slope of exactly −1 places it exactly there. Below it the system would freeze; above it, thermodynamic quantities would diverge.
Fit a maximum-entropy model to the measured firing rates and pairwise correlations, then multiply every fitted parameter by 1/T and sweep T. This is a purely fictitious temperature: it does not correspond to heating anything, it just reweights the model’s configurations. Compute the specific heat,
The evidence is not the peak. It is the trend: a peak sitting at T = 1 that grows and sharpens as you include more units is the finite-size signature of a genuine divergence. In salamander retina the peak height climbs monotonically from roughly 4 at N = 5 to roughly 26 at N = 40, with a simulated N = 120 network near 49. The rank–frequency plots likewise approach the −1 line more closely as N grows.
Beggs & Plenz recorded spontaneous activity from cortical cultures and slices on a 60-electrode array and defined an avalanche as a run of consecutive non-empty time bins.6 The numbers are worth carrying:
That last point is the strongest structural feature of the affirmative case: there is a knob, and turning it destroys the signature in the predicted direction. A power law that cannot be broken is a description; one that breaks when you push the system off its operating point is closer to a measurement.
Muñoz’s verdict on his own field is: “for most of the leading examples… there is not a fully convincing proof… we still do not have a ‘smoking gun’.” His Figure 1 — the first figure of the review — is a plain unbiased random walk producing clean power laws with exponents τ = 4/3 and α = 3/2 and no critical point anywhere. He states outright that finite systems cannot be truly critical, and that externally driven systems are not truly critical either — for those, the best available surrogate is the Widom linePast a critical point, the sharp transition is gone but the locus of maximum response or maximum correlation continues into the supercritical region. It is a ridge, not a point, and it is what a driven or finite system can actually sit on., the locus of maximal susceptibility. He also records the tautology objection against himself: fitting a rich dataset tends to return a near-critical model because that is the only region where the model can generate many scales at once.
The negative literature has not disproved criticality in living systems. What it has done is raise the evidentiary bar, and it has done so on every kind of evidence at once. The single most important result is the one below: both signatures of Act II arise generically from a mechanism with no criticality, no coupling and no tuning in it at all.
Take N binary units. Suppose they are conditionally independent given one unobserved variable h — a common input, a stimulus, an arousal level, a driver. Conditional on h, no unit knows about any other. There is no coupling between units whatsoever. What you actually measure is the average over h:
Now follow the algebra in this order. It is four steps and it is worth doing slowly.
A latent variable sweeps the population through a range of rates; marginalising mixes conditional distributions of different entropies, and that mixture alone manufactures the power law. Nothing is poised — something is merely being driven.
Schwab, Nemenman & Mehta proved this in five pages.24 Their numbers are the ones to quote. Fitting the rank–frequency plot of N = 200 non-identical conditionally-independent units gives a slope of −1.012. Repeating with genuine quenched spin–spin interactions — conditional independence deliberately broken — gives −1.011. The effect is robust over two orders of magnitude in the width of the hidden variable’s distribution; it fails only when that width is so small there is nothing to average over. And on real data — a single motion-sensitive blowfly neuron — a manifestly non-critical refractory-Poisson surrogate driven by the same stimulus reproduces the observed Zipf plot.
Aitchison, Corradi & Latham proved the mixture version in general and, importantly, supplied a diagnostic that can come back negative: the proportion of energy variance explained by the candidate latent variable.25 Applied to zebrafish antibody repertoires it returned below 0.5 in 12 of 14 datasets, and Zipf’s law held at every individual sequence length — so for that dataset the mechanism is not the explanation. A test that can fail, and did.
Everything above is one toggle. Sixteen units. Each flips on with a probability that depends on a hidden driver h and on its own fixed bias. There are no connections between the units — look at the left panel and satisfy yourself that there is nothing linking the dots. States accumulate into a rank–frequency count in the middle panel, and the same counts appear on log–log axes on the right, against a slope −1 guide.
With the latent variable ON, h drifts and the log–log points settle onto the guide line. With it OFF, h is frozen at its mean — the units are still random, still noisy, still exactly as uncoupled — and the plot is visibly not a straight line. That single toggle is the whole teaching point.
Toggling restarts the sampler, because the two conditions must not be mixed. The fit runs over ranks 2–400 and settles after roughly forty thousand samples. Expect about −1.0 with the driver on, and about −0.4, shallower and visibly bent away from the guide, with it off. The gentle droop at the far right in both conditions is undersampling of rare states, not physics — it is the same finite-size cutoff every real avalanche distribution has.
If you want to claim criticality in a system that has an external driver, these are the four things that separate your claim from the null above. They are listed strongest first.
The closest published analogue to a slowly driven, sparsely instrumented natural system is Priesemann & Shriki 2018, and it is worth reading end to end.26 The setup: completely unconnected Poisson units, no interactions, no propagation, no transition — but the rate at which they fire wanders slowly, over one or two orders of magnitude, on timescales long compared with the events being counted.
A homogeneous Poisson process cannot make a power law: its event-duration distribution is exactly exponential. But under a slowly varying drive each window behaves like its own homogeneous process with its own decay constant, and the pooled distribution is a mixture of exponentials:
Two results from that paper are the ones that bite:
The paper cites solar flares explicitly as the textbook case of heavy-tailed statistics generated by superposing exponentials at different event rates. Anyone applying criticality machinery to the magnetosphere — an externally driven system whose driver varies over orders of magnitude on timescales far longer than the events being counted, observed through a sparse and unevenly spaced ground network — is in exactly this neighbourhood, and inherits the whole objection. The authors also give the seven-measure battery that separates a driven Poisson mixture from a genuinely critical branching process, of which the sharpest single item is the mean-size-given-duration relation: the driven null gives the trivial exponent 1, criticality requires the value implied by the marginal exponents.
And they state the honest limit themselves, twice: a sufficiently sophisticated drive can reproduce anything a passive analysis measures, so only causal intervention can decide. For a system nobody can intervene on, the correct posture is to state the drive-only null explicitly, report every discriminator, and frame the result as “not explained by the drive-only null under tests 1–4” — never as “the system is critical.”
Six independent lines, each attacking a different kind of evidence.
Morrell, Sederberg & Nemenman reproduce the full coarse-graining scaling signature — activity variance α = 1.367 ± 0.009, free energy 0.841 ± 0.009, eigenvalue spectrum −0.648 ± 0.009, correlation time 0.273 ± 0.010, plus flow to a non-Gaussian fixed point — from neurons with zero cell-to-cell coupling, matching published experimental values within error.27 The one ingredient that matters is that nearly all cells couple to latent fields, and that there are at least five of them; earlier single-latent-variable models failed for exactly that reason. The 2024 sequel extends the construction to avalanche criticality.28
The accepted gold standard was never a power law alone but the full package: power-law sizes and durations with the right exponents plus collapse of average event shapes onto a universal scaling function. Touboul & Destexhe obtain the entire package from N = 2000 completely independent Poisson processes sharing one fluctuating rate — size exponent 1.47, duration 1.9, and an accurate shape collapse.29 Their earlier paper had already shown that thresholding a stochastic process yields power-law avalanches.30 Mariani et al. go further: with all coupling set to zero, a shared slow modulation reproduces the power laws, the crackling scaling relation and the shape collapse — but it pins the correlation length at 1, which is precisely why correlation-length-versus-system-size is the discriminator that survives.31
Observe a spreading process through a small subset of its units and the conventional branching-parameter estimator is badly biased. Wilting & Priesemann give a subsampling-invariant estimator; applied to rat, cat and monkey it returns a reproducibly subcritical cortex, median m̂ = 0.984 — neither Poisson-like nor critical, which they name the reverberating regime.32 The same cat data read 0.27 under the conventional estimator, and 0.06 from single neurons: it would have classified cortex as asynchronous-irregular. The worst case for a fixed array is older and starker — one and the same critical model reads “apparently supercritical” at small electrode spacing and “apparently subcritical” at large.33 The constructive fix is subsampling scaling.34 Their review of the whole dispute is the fastest way in.35
Bédard, Krüger & Destexhe recorded global and single-unit activity simultaneously: the units were Poisson, the avalanches exponential, and the global 1/f spectrum was supplied by the filtering properties of the extracellular medium.36 The scale-free behaviour was a property of the measurement chain, not of the network.
Clauset, Shalizi & Newman supply the maximum-likelihood + Kolmogorov–Smirnov + likelihood-ratio standard, and then apply it to 24 canonical datasets.37 Exactly one — word frequencies in English — earns the rating “good”. Seven are firmly ruled out. The log-normal is not excluded for any dataset but one. Least-squares fitting on a log–log histogram, still the commonest method in practice, returns exponents off by around a full unit, with error bars that give no warning of the bias. Gerlach & Altmann then show that recipe assumes independent samples and breaks under correlation: an earthquake series of N ≈ 105 carried about 75 effective independent samples.38 And Destexhe & Touboul report that 100% of 32,000 non-critical distributions passed an information-criterion comparison of power law versus log-normal — the test does not discriminate.39 The short, sharp summary of the whole problem is Stumpf & Porter.40
The Fisher information of a model equals its susceptibility. Since susceptibility diverges at a critical point, maximum-likelihood inference on high-dimensional data preferentially returns models near a phase transition — there are simply more distinguishable models there.41 The consequence is developed further in later work.4243 Related routes: random subsampling alone generates signatures of criticality in simulated networks,44 neutral dynamics generates scale-free avalanches with no critical point,45 non-critical neuronal systems produce power laws by other routes,46 and the extrinsic-versus-intrinsic distinction has since been formalised directly.47
It comes from the discoverer of neuronal avalanches, who wrote a paper cataloguing the objections against his own result as a dialogue.48 He concedes the central point in one line: “A critical system will produce power laws, yes, but power laws do not prove criticality.” He then gives the three things that do go beyond power laws, and they remain the best checklist in the literature: (1) tunability through phases — a control parameter that sweeps the system from sub- to supercritical, with non-power-law distributions on either side, which the non-critical generators simply cannot do because they have no phases; (2) mathematical relations among several exponents, which single-exponent generators cannot have; (3) data collapse of avalanche shapes. Note that the deflation literature has since partly caught up with (3), which is why (1) and (2) now carry the weight.
What the negative literature has actually accomplished is this: “a power law was observed” is now worth nothing as evidence, and the minimum admissible claim is a rejectable exponent relation plus a null model that fails. The question of whether cortex is critical remains open.49
Amid all of this, one component of the original Kauffman programme was taken up, made rigorous, and turned into a working research field with theorems and empirical applications. It is also the component with the least contact with everything else on this page.
Kauffman’s 1986 argument is a random-graph connectivity argument: in a soup of polymers where molecules can catalyse each other’s formation, once the catalysis probability crosses a threshold, a collectively autocatalytic set — a set of reactions in which every reaction is catalysed by something the set itself makes — appears almost inevitably.2 It is an appealing origin-of-life story. Its problem was that “does this system contain such a set?” looked like a search over exponentially many subsets of reactions, and the claim itself was disputed — the counter-position being that if the catalysis probability scales inversely with the number of reactions, such sets may never appear at all.
They turned the conjecture into a decidable object.50 Define a catalytic reaction system as a triple — molecules, reactions, and a set of catalysis pairs — over a distinguished food set that the environment supplies freely. Then a subset of reactions is a RAF — reflexively autocatalytic and food-generated — if:
The algorithm avoids the exponential search entirely by shrinking rather than searching. Two reduction rules — delete any reaction with no available catalyst; delete any reaction whose reactants cannot be built from the food set — are applied alternately until nothing more can be removed:
The theorem: a RAF exists if and only if R∞ is non-empty, in which case R∞ is the unique maximal RAF. It is computed in polynomial time — a worst case of O(|X||R|3), measured empirically at τ = 0.0002 |R|1.43 — and an irreducible RAF, if one exists, in O(|X||R|4). That is the whole difference between a suggestive story and a research programme: the object became something you can compute, and therefore something you can be wrong about.
This is the component that got a rigorous treatment — alone, and by people who touch neither topology nor criticality. In a fifteen-paper sample of the RAF literature, the vocabulary of the other two senses does not appear at all: no “poised realm”, no level statistics, no avalanche exponents. It is a working field that happens to descend from the same author. It is not evidence for anything else on this page, and it does not claim to be.
This is where the evidence is unambiguous and the finding is negative in a specific and unusual way.
Space the energy levels of a disordered quantum system and look at the gaps between neighbours. Three shapes are possible, and they are the fingerprints of three regimes:
The proposal: biomolecules sit at that critical point, without external tuning. Below it, coherence lasts but nothing can travel; above it, transport works but coherence dies at the rate the warm environment sets. Only at the transition do long coherence times and extended, transport-capable states coexist — and biology, the argument goes, lives there. Kauffman’s framing adds that incoming photons repeatedly re-purify the system, letting it “hover” reversibly in a Poised Realm between quantum and classical.
This is easy to miss and important. The 2014 PLoS ONE paper9 analyses no molecules. It builds no Hamiltonian from any molecular structure. It never computes, histograms or fits a level-spacing distribution — the three formulae above appear only as cited background. The object simulated is an abstract 25-site one-dimensional model ring with quasi-periodic on-site energies. The physics chain it lays out is standard and correctly cited, and its numerics do demonstrate that the mechanism exists in a toy model.
Its single empirical claim is one figure: a decay exponent of about −0.23 compared with a theoretical −1/4. That agreement is a by-eye match between two hand-drawn guide lines of identical slope on log–log axes whose time and amplitude units are arbitrary for both traces, over roughly one to one-and-a-half decades of data that oscillates by up to a full log unit about the trend, with no error bar — and the only real-world input is one digitised population-decay curve for one light-harvesting complex, obtained from another group. No off-critical comparison curve is plotted on the same axes, so a reader cannot see whether a non-critical model would have looked different.
The molecule analysis is in the 2015 DICE proceedings companion.10 Its evidential structure, item by item:
| Ingredient | What the paper actually has |
|---|---|
| Molecules | 18 with displayed results, of which 9 are assigned “critical”. A further 15 small molecules are asserted with no displayed data, and all 20 amino acids are classified — the paper’s own words — “by inspection”, with the conclusion stated as what “is likely”. The multifractal analysis is shown for one molecule. |
| Level of theory | Not density-functional theory. Semi-empirical extended Hückel, valence orbitals only, one global constant. The parameter set is never specified. The whole result is a property of that matrix, and it is never validated against a higher level of theory. |
| Unfolding | A local running average whose window is given only as a range, “2…5”. Which value produced the figures is never stated, and no sensitivity analysis is reported. |
| Non-biological control | None. Their own data crosses the biological line anyway — silk and DNA read Poisson, steroids read Wigner — and this is handled by a post-hoc re-partition of biology into “active biochemical”, “structural” and “polycyclic”. |
| Statistical test | None. The entire quantitative support is “less than 3% error” — a maximum deviation. No p-value, no critical value, no likelihood ratio, no confidence interval. At their own 6,328 level spacings a naive 5% Kolmogorov–Smirnov critical value would be about 0.017 — below the paper’s own 3% bound. |
| Null model | None. No surrogate spectra, no randomised Hamiltonians, no scrambled geometries, no size-matched synthetic organics. The paper’s strongest rhetorical claim — that hitting the critical point by chance is “astronomically” improbable — is asserted with no ensemble defined and no probability computed. |
| Uncertainties | Exactly one in the whole paper, and it is not on the criticality claim: a formal slope error on a straight line fitted to a hand-chosen scaling window. |
The paper presents the absence of fitting as a strength — “no parameter fitting is involved in the procedure”. It is not. Zero free parameters removes the need to penalise for them; it does not supply a rejection criterion. A zero-parameter comparison with no null distribution has no rejection region at all.
Not refuted — unengaged. And it was never refuted partly because, as written, there is no statistic in it to fail. Both halves matter, and they are different findings. The first is about the claim’s audience; the second is about its construction, and it is the one that should govern any future use of it.
Why unengaged is a worse outcome for an idea than being refuted. A refuted claim leaves something behind: a measurement, a method, a bound, a reason. Someone ran the experiment, and the field is better calibrated afterwards. An unengaged claim leaves nothing — and it accumulates a citation trail that looks like uptake from a distance, so it can be mistaken for support by anyone who counts citations without reading them. Worse, the absence of rebuttal starts to be offered as though it were evidence. It is not: a claim with no rejection region cannot be refuted by anybody, so silence about it is uninformative.
The claim is not dead. It is untested, and it is cheap to test — dense diagonalisation at this size is a laptop job now. The decisive missing experiment is a single number: what fraction of arbitrary size-matched organic molecules also land on semi-Poisson at the same tolerance? If the answer is “most of them”, the claim is about large disordered molecules and not about life, and the evolutionary framing collapses without anything in the paper being wrong.
On Yamanaka 2017. The full text could not be obtained — the publisher returns an automated-client block on a paper that two independent services report as open access. Its characterisation above therefore rests on indexing metadata and abstract-level evidence, not on a reading of the paper. It is the single most consequential gap in this account, because every claim about whether the poised-realm hypothesis was ever independently tested rests on that one paper’s contents.
On the trajectory. The author’s sole solo posting in this orbit is a philosophical extension to mind–body and free will, with no level-statistics content whatsoever.64 Later work from the original group continues in the physics direction.65 These are reception facts about where the idea went, not judgements about anyone.
| Sense of “poised” | Studied by whom | Principal critique | Current status | Verdict |
|---|---|---|---|---|
| (a) Edge of chaos CA / Boolean networks |
Langton, Packard; re-run by Mitchell, Hraber & Crutchfield; modern re-tests in reservoir computing and neuromorphic hardware | The launching experiment’s clustering was an artifact of task symmetry, combinatorial drift and search symmetry-breaking — a no-selection control reproduces it | Survives as a substrate-specific heuristic; not as a law | Deflated |
| (a′) Autocatalytic sets RAF theory |
Kauffman; formalised by Hordijk & Steel; structure theory and metabolic application by Steel, Hordijk & Xavier | None material. Boundaries are stated honestly (smallest RAF is NP-hard; irreducible RAFs can grow exponentially) | Working field: theorems, polynomial-time algorithms, application to reconstructed ancestral metabolism | Formalised |
| (b) Poised at criticality living systems |
Beggs & Plenz, Mora & Bialek, Cavagna, Tkačik, Muñoz — against Schwab, Aitchison, Touboul, Priesemann, Wilting, Morrell, Marsili, Clauset | Latent variables, external drive, subsampling, thresholding, binning, pooling and the inference procedure itself each reproduce the signatures without criticality | Genuinely open. The bar has moved from “a power law was observed” to “a rejectable exponent relation, plus a null model that fails” | Contested |
| (c) The Poised Realm quantum ↔ classical |
Vattay, Kauffman & Niiranen and the 2015 companion. One independent technical attempt, itself cited zero times | No control set, no statistical test, no null model, one stated uncertainty; the flagship computes no level statistics at all | Untested. Not refuted — and no rejectable statistic was ever stated | Unengaged |
No. On the evidence assembled here, no source studies the three senses as one object, under this or any other name. Four observations support that, and each is checkable:
This is what a literature search found, not a proof of absence. The evidence base is a 77-paper corpus assembled around five named clusters, with verified identifiers and full-text extractions for the load-bearing items — not an exhaustive search of the literature. It is strong enough to say that no prominent unified treatment exists in this literature. It is not strong enough to say that none exists anywhere. One specific gap is known and named: a proper forward-citation study of The Origins of Order has not been done, and would be the obvious next step.
It would be easy to read everything above as a counsel of despair. It is not. There is a published standard that shows what the affirmative case looks like when it is done in a form that could fail, and it is worth studying more closely than any of the deflations.
Fontenele et al. do not merely fit power laws.66 They test the crackling-noise scaling relation:
The two sides come from three separately estimated exponents, from three separate fits, which are never constrained to agree. So the equality can fail — and it does fail almost everywhere. Rather than treating a recording as critical or not, they parameterise cortical state continuously by spiking variability and sweep it. The relation holds only in a narrow band, at intermediate variability, and at the same place an entirely independent signature — long-range temporal correlations — peaks. Exponents at that band: τ = 1.52 ± 0.09 and τt = 1.7 ± 0.1 in rat, at ⟨CV⟩* = 1.4 ± 0.2; in freely moving mice, ⟨CV⟩* = 1.1 ± 0.2. And the second exponent falsifies the textbook universality class: τ matches mean-field directed percolation, τt does not.
That is the form to imitate, and the reason is structural rather than statistical: a power-law fit cannot fail, because any heavy tail yields some exponent. A relation between three independently measured exponents can. Two caveats keep it honest, and both should be carried forward: the paper asserts the agreement graphically, via overlapping standard-deviation bands, and never quotes a numerical residual at the crossing;66 and Destexhe & Touboul showed that non-critical models also satisfy the same relation within a connected region of avalanche-cutoff choices, so the criterion is only as good as a cutoff policy fixed in advance.39 The lesson is not “use this test”. It is: state, before you look, the number that would have made you abandon the claim.
Ordered for someone who has read none of these. Every entry gives the overview first, then why you should care, then difficulty and rough reading time, then the source. Difficulty is relative to a reader comfortable with undergraduate statistical mechanics and probability, not to a specialist. Times are for a first careful pass, not a skim and not a re-read.
A review that fixes what criticality is and is not, walks the evidence at every biological scale from neurons to bird flocks to gene networks, and states the objections to each case as it goes. It is the only document in this set that genuinely spans two of the three senses of “poised”.
Why you care: it opens with a null model. Figure 1 is an unbiased random walk producing clean power laws with the “critical” exponents and no critical point anywhere — the field’s own authoritative synthesis leading with a warning against the inference it then defends. Read Sections I–III properly and Section IV selectively; Appendices A (generic scale invariance) and B (statistical criticality) are short and are where the sharpest caveats live.
A careful re-run of the experiment that launched “evolution adapts toward the edge of chaos”, finding that the reported clustering is produced by three mechanisms that have nothing to do with criticality — and demonstrating one of them with a control run in which fitness is never evaluated at all.
Why you care: this is the single most transferable methodological lesson on this page, and it is almost free of mathematics. If you read only the figures — the original histogram, the replication histogram, and the pair of no-selection / with-selection mosaics — you will have the argument. The takeaway generalises well beyond cellular automata: a peak in the distribution of a control parameter over an evolved ensemble is not evidence of a critical point, because the measure on the underlying space can produce it unaided.
The proof that Zipf’s law — and with it the whole “statistical criticality” inference — arises generically whenever the observed units are coupled to one unobserved fluctuating variable. No coupling between units, no parameter tuned to anything.
Why you care: it is section 4 of this page in the authors’ own words, and it is short. If the saddle point and the Gärtner–Ellis argument are heavy going, read only the paragraph beginning “If the extensive parts of the energy and entropy do not cancel…” and Figures 2 and 3 — the physical statement and the two numbers, −1.012 and −1.011, take half an hour. The second of those numbers is with genuine interactions switched on, which is the point.
The paper that popularised the modern hypothesis. It builds statistical mechanics directly from data across four unrelated systems — retina, protein families, antibody repertoires, bird flocks — and argues that the fitted models keep coming out at a special point in their own parameter space.
Why you care: read it after item 3, so that you can watch the argument being constructed with the objection already in mind — the experience is much sharper in that order. Its list of ten operational tests, scattered through the text, is the checklist a referee will hold you to. It is also unusually honest about its own limits: the authors call criticality “a hypothesis whose most compelling tests are yet to come” and note that the inference problem is hardest exactly near criticality, which is the regime being claimed.
Spontaneous activity in cortical tissue on a 60-electrode array, parsed into cascades. Size exponent −1.50 ± 0.008 over about 90,000 events; branching parameter measured directly at 1.04 ± 0.19; and a pharmacological knob that destroys the power law and then restores it on washout.
Why you care: this is the measurement everything in this field argues about, and it is a genuinely well-executed experiment — note the array-size control that shows the cutoff is instrumental, and the jitter control that shows the branching parameter is not a chance property of the firing rate. Read it for what a careful version of this claim looks like, and for the fact that two independently computed quantities cross at the critical values at the same bin width.
Instead of asking “is this recording critical?”, it parameterises cortical state continuously and asks where along that axis a relation between three independently measured exponents is satisfied. The answer is: in a narrow band, and nowhere else.
Why you care: this is the one paper here that a reader should take away as a template. It is short, it is dense, and it demonstrates the difference between a description and a prediction: a power-law fit cannot fail, an exponent relation can. Read it alongside its own critic (item 12) — the two together teach the real lesson, which is about pre-registering your cutoffs, not about this particular criterion.
Unconnected Poisson units under a slowly wandering rate produce a mixture of exponentials that mimics a power law with the canonical critical exponents. Fully analytic, and it names the discriminators that separate the mixture from a real critical process.
Why you care: it is the closest published analogue to any externally driven natural system, and it explicitly cites solar flares as the classic case. Two things are worth the price on their own: that fixing the duration exponent at the critical value delivers the size exponent for free, so recovering both is not independent evidence; and Figure 6, where pooling twelve real recordings with different mean rates converts twelve non-power-laws into one clean power law. If the mixture integral in Section 5 is heavy, read Section 5’s conclusion, the seven-measure table, and Figure 6.
The maximum-likelihood + Kolmogorov–Smirnov + likelihood-ratio procedure that turned “it looks straight on log–log axes” into something testable, applied to 24 canonical datasets.
Why you care: exactly one of those 24 — word frequencies in English — earns the rating “good”. The log-normal cannot be excluded for any dataset but one. Do not read this front to back; read Sections 3–5 for the method, then the two verdict tables in Section 6, and keep it open when anyone shows you a straight line. Its most useful single table shows that fitting a straight line to a log–log histogram — still the commonest method in practice — returns exponents off by around a full unit, with error bars that give no warning.
The paper that turned autocatalytic closure from a suggestive story into a decidable object with a polynomial-time algorithm.
Why you care: it is discrete mathematics, not physics, and it is a complete change of register after items 1–8 — refreshing rather than taxing. Read Sections 2–3 for the definitions, Section 5 for the two reduction rules, and Theorem 6.1. The lesson that transfers: the field became productive at the exact moment the central object became computable, and therefore something one could be wrong about.
The first obtains power-law sizes, power-law durations and a universal shape collapse from 2000 completely independent Poisson processes sharing one fluctuating rate. The second does it with all coupling set to zero and gets the crackling relation too — but its construction pins the correlation length at 1, which is exactly why correlation length versus system size is the discriminator that survives. Read the second for that contrast; it is the cleanest statement of which signatures track coupling and which do not.
Four coarse-graining exponents plus flow to a non-Gaussian fixed point, all reproduced within experimental error by neurons with zero cell-to-cell coupling. Note the condition that makes it work — at least five latent fields, coupled to nearly every cell — because it explains why earlier single-common-input models failed and is the thing to test for.
The first: observing a spreading process through a small subset of its units biases the standard estimator badly, and the corrected estimator returns a reproducibly subcritical answer. The second is the direct critique of item 6, and shows an information-criterion test passing 32,000 out of 32,000 non-critical distributions. Read them together as “what your instrument does to your answer”.
Structure theory first — the whole lattice of autocatalytic subsets characterised for a biochemically motivated restriction, with the hardness boundaries stated plainly — then the application to reconstructed ancestral metabolic networks, which is the answer to “this only ever described random graphs”. A gentler entry point is the book-length introduction.57
Read them in that order, and read the second for the evidence. Go into the first expecting a proposal and a toy-model demonstration, not a level-statistics analysis — there is none in it, and the whole empirical content is one borrowed decay curve. Then read the companion with four questions in hand: how many molecules, is there a non-biological control, is there a statistical test, is there a null model. The exercise takes under two hours and is the fastest way to calibrate what “unengaged” means in practice.
The first is written as a dialogue between a believer and a sceptic by the discoverer of neuronal avalanches, and it is the most enjoyable item on this list. The second gives a diagnostic for the latent-variable mechanism that can return a negative verdict — and does, on one of its own datasets, which is why it is worth having.
The solo philosophical extension of the Poised Realm to mind–body and free will contains no level-statistics content, and reading it will not tell you anything about whether the physical claim is true.64 And the 2015 proceedings paper that looks like independent uptake of that claim is a self-declared review with no original calculations, published in the same conference volume — it is co-location, not confirmation.60 Both are recorded here so the elimination does not have to be repeated.
Muñoz’s Figure 1, Mitchell’s no-selection control, and the toggle in section 4. Between them they say the same thing three ways: before you can claim a system is poised, you have to state what would have shown it was merely being driven.
Every entry below corresponds to a verified bibliographic record. Where a work genuinely has no DOI — the 1988 book chapter, the 1993 and 1994 Complex Systems / SFI papers, and two arXiv-only postings — that is stated and the arXiv identifier or venue is given instead. No DOI has been supplied by inference.